Hey @fede (nixtla) (they/them), how do you manage input

DataFrames

to fit the model? From what I see it is transformed into ndarray that is later processed, however, couldn't find the function that transforms from DataFrame into np array. Wanted to add polars support

Hi, how do I reuse the model for new data? So literally storing the fitted and optimised model and using predict for some new observations (I see the new_data argument in .predict but I need a fitted model first).

Hi I was trying to use AutoARIMA with Exogenous Regressors (https://nixtla.github.io/statsforecast/examples/exogenous.html) for my current ts problem. 1. Is it a normal limitation for the AutoARIMA model that I need to know all the future regressors (it seems to me that I can only use obvious features like "is_holiday" or "weekday_sin/cos" then)? 2. I would like to compare it to my cross-validation runs. Are you planning to implement this for exo regressions as well?

Hi @fede I have a question about https://nixtla.github.io/statsforecast/models.html#simpleexponentialsmoothingoptimized. Is this function a replica of HoltWinters in R or somewhat different. Is there a way to access optimized alpha from the function run like in R? I am trying to run HoltWinters in R on the same data to match the R output, is sesopt the best choice to match R output?

Hi guys, a quick question/suggestion. In

mlforecast

in

fcst.preprocess()

, the default value for

dropna

is True, so in cases when we have intermittent series and we generate features like

rolling_std

, NAs might occur anywhere in data (not just at initial rows) and we are dropping all those observations. the consequence might be inaccurate CV estimate when we try to evaluate the model. May be a warning in the function to notify the user about the default value of

dropna

?

Hi everyone, Is there any way to get fitted values in statsforecast spark pipeline (FugueBackend), right now there are only two functions forecast and cross_validation_fitted_values

hi everyone! is it possible to call the new

.forward()

method with a single timestamp? we have series with high autocorrelation, and would like to train a simple AR model (or ARIMA), but would like to update the model on each timestamp, so it’ll have access to T-1. when predicting multi-step ahead (calling .forward() in a mini-batch fashion), the ARIMA model underperforms our naive model, and we’d like to benchmark these two against each other! we’re writing a TS cross-validation library, with the ability to also online update models on each timestamp, this is the PR with the failing tests we have for

statsforecast

wrappers, we’d love to put to use! 🙂 https://github.com/dream-faster/fold-models/pull/38

Hello guys,

Hello guys, how can I make prediction on training data, I am using mlforecast all I can see is how to predict on the next steps using horizon but what if I want to predict on my training data as well please suggest some steps

Hi Guys! Q: Did anybody implemented a way to integrate AutoArimaProphet directly in a StatsForecast object ? Like in that

models

list variable of the StatsForecast object, to be able to also add AutoArimaProphet & then use the

.forecast()

for all the models at once (again, including AutoArimaProphet)

Hello all! I've got a question: is it possible to add future regressors to an AutoARIMAProphet model? Do I need to add them in the arima model? Or do I just need to change the disable_seasonal_features from

True

to

False

?

Hello Guys,

Hello Guys, I need to do some hyper parameter tunning with MLforecast the data I am using is very large and it is becoming computionally expensive can. you please suggest some resources for light weight hyper parameter tunning.

Greetings! It is a pleasure to be here. I became acquainted with Nixtla through a recent LinkedIn post by Max, and I must say that I am thoroughly impressed. The creation of a dedicated ecosystem for time series modelling is an invaluable resource, and I have already begun to utilize it. After thoroughly exploring the Nixtla GitHub repositories, I discovered state-of-the-art solutions for time series forecasting. I am curious, does your company have plans to extend the ecosystem to include other tasks such as time series classification, unsupervised time series clustering, and the like? These fields are increasingly popular, yet there are currently few frameworks that can handle time series data for these tasks, and the most widely used method of employing RNN/LSTMs is computationally expensive. If your team is planning to delve into these areas, I am eager to contribute to both the research and implementation aspects in an open-source capacity. Currently, the only existing solution for these tasks is the sktime package in Python. While similar to sklearn, it is less intuitive to use and may not yield the most accurate results. cc: @Max (Nixtla) @fede (nixtla) (they/them)

Hello everyone, I have come to know about nixtla from databricks support team while our company was browsing for solutions for distributed forecasting. Before we move forward with implementation of nixtla we want to verify that nixtla has following features : 1. Distributed computing across cluster and workers in spark environment 2. Compatible in GPUs and Multi node GPU 3. Bring your own models 4. Control over dataset creation Could this community provide blogs or links or answers this questions Cc : <!here> @Nixtla Team

Hello everyone, I have just started using nixtla for hierarchical timeseries forecasting for a class project. I have been trying to replicate the example on the “TourismSmall” dataset using the ARIMA model and the MinTrace(method=‘mint_cov’) as a reconciler. However, I always get the following exception: Exception: min_trace (mint_cov) needs covariance matrix to be positive definite. I also tried other models (for example, AutoETS), other datasets (for example, Labour) and I also played with the mint_shr_ridge parameter but I still get the same error. Could you let me know how I can overcome this issue? Thank you very much in advance! Asterios

Hi all, is there a way on neuralforecast to use multiple CPUs/GPUs for cross-validation if I only have a single time series?

Hi all, I am using statsforecast library while including multiple exogenous variables. Are there any best practices to analyze the impacts/importance of these features and to select the most efficient set of features?

Are there particular cases where you think the mlforecast library approach might offer better predictions than the algorithms in the neuralforecast library?

Hi All, I recently came across neuralforecast and would like to implement it in my usecase, where I have to predict next basket size of a customer, so I have prepared the dataset at batch level for each customer. And, each batch is having day wise purchases of a customer, recency, interval, frequency, avgsize of his basket, etc. for 5 days window and based on this sequential purchases, I need to predict his next purchase which is my y_batch..Can anyone please help me how can I use neuralforecast as I have multiple features i.e. 13 features in my batch. Please let me know in case of any questions. Thanks in advance!

Hi, I wonder if mlforecast supports categorical static features?

Does anyone know what maybe the issue here? I couldn’t show the data due to confidentiality but all the static features are numeric and only id, date, y are left.

Hi team, I am having some issue with the statsforecast models. I have numeric y (validated to be all positive values) and other 2 numeric variables. I am using Naive, RWD, AutoARIMA, AutoETS, AutoCES and AutoTheta models. I am constantly getting the warning messages in the 1st screenshot, which seems to be related to ARIMA model and ETS model. Plus, I tested removing ARIMA and ETS model from the list, however, I am getting the error message in 2nd screenshot. Does anyone know what might be the cause for any of these errors?

Hello Team,

Hello Team, I am using MLforecast one of my projects where we are predicting demand for 20-30 large datasets I am currently using pandas udf with pyspark to implement MLforecast, However my code breaks down while applying MLforecast on one of the large datasets. I want to implement MLforecast using Dask or some other library can anyone please help me with any examples of Dask implemenations of MLforecast. I am stuck with this I request members to help

Hello team, Quick question: the newest version of neuralforecast has a runtime dependency on Pytorch 2.0.0 (and the corresponding lightning package). Is this necessary, or can I still use 1.13? There are no tests to verify if my installation will work with an older version like that. (Tbh I only want maybe 1.12 or 1.13 support, nothing older).

Interesting to see Google is putting a new model into production before large-scale benchmarking https://ai.googleblog.com/2023/04/recent-advances-in-deep-long-horizon.html?m=1

Hello! Thank you Max and Federico for teaching me about conformal prediction today, I think the work you guys do is fascinating. I would love to get more involved and contribute to the open source code- are there any good tasks/resources for first time contributors?

^Agreed, today's talk at PyData Seattle about mlforecast / conformal prediction was fascinating! Would you be able to share the slides you presented? (Thanks!)

Welcome @kaixu yang! Dear community, we are very happy to have Kaixu here. He is one of the main developers of GreyKite. https://github.com/linkedin/greykite Hopefully we can learn something from him.